# RECORD ddfcdefd81b294b9edfe4b97744d2d4725aeed23 # TITLE: Urban location and housing prices within a hedonic model # YEAR: 2008.0 VENUE: Journal of Regional Analysis & Policy DOI: 10.68479/001c.9464 # AUTHORS: J. Ottensmann; Seth B. Payton; J. Man # PAGES: 17 CANDIDATE TABLE PAGES (1-based): [4, 8, 9, 12, 14] # ABSTRACT: A measure of location relative to employment is often included in hedonic housing price models. This is most often distance to the center, based on the monocentric model, which does not consider the decentralization of employment in urban areas. This paper tests the performance of alternative measures of location, considering both distance and time to the center and to multiple employment centers and measures of accessibility to employment and change in accessibility. The measures using multiple employment centers and accessibility perform better than simple distance to the center, with the combination of accessibility to employment and change in accessibility doing best. ===== PAGE 1 (score 6, numeric-paren 0, table-words 0, var-words 6) ===== Urban Location and Housing Prices within a Hedonic Model John R. Ottensmanna, Seth Paytona, and Joyce Manb Indiana University–Purdue University Indianapolis a and Indiana University Bloomington b - USA Abstract. A measure of location relative to employment is often included in hedonic housing price models. This is most often distance to the center, based on the monocentric model, which does not consider the decentralization of employment in urban areas. This paper tests the perfor- mance of alternative measures of location, considering both distance and time to the center and to multiple employment centers and measures of accessibility to employment and change in accessibility. The measures using multiple employment centers and accessibility perform better than simple distance to the center, with the combination of accessibility to employment and change in accessibility doing best. 1. Introduction Location within an urban area, particularly with respect to the location of employment, is assumed to be a determinant of land prices within standard urban economic models. And because land prices, ceteris pa- ribus, will affect house prices, one or more measures of urban location are frequently included in hedonic housing price models. The standard urban economic model (Alonso 1964; Mills 1972; Muth 1969) makes the monocentric assumption that employment is located in the Central Business District (CBD). Following this, distance to the CBD has frequently been used as a measure of loca- tion. With the continuing movement of employment in metropolitan areas away from the CBD, distances to multiple employment centers or even to all employ- ment in the form of an accessibility measure is increa- singly used as an alternative. This paper examines the performance of alterna- tive measures of urban location within a hedonic housing price model for Indianapolis, Indiana. The goals of the research are to understand better how res- idence location relative to employment location affects house prices and to provide guidance as to the most effective method for specifying location within hedon- ic models. The research systematically compares measures of location using both distances and travel times, to the CBD, to multiple employment centers, and used in measures of accessibility to employment. The significance of changes in accessibility to em- ployment is also assessed. 2. Location in Hedonic Models In their seminal research developing a hedonic model for the prediction of house prices, Kain and Quigley (1970) included structural characteristics of the housing unit, neighborhood characteristics, and distance to the CBD. Reviews of hedonic models (Bo- wen et al. 2001; Malpezzi 2003) have described such models as including housing structure characteristics, the social and natural environment (neighborhood characteristics), and location within the market. Heikkila et al. (1989) cited numbers of studies of determinants of residential property or land values using hedonic models, saying that to the extent they have included location, it has generally been distance to the CBD, using the assumption of a monocentric city employed in the standard urban economic model (Alonso 1964; Mills 1972; Muth 1969). The early model developed by Kain and Quigley (1970) included dis- tance to CBD, but it was not statistically significant. Witte et al. (1979), in one of the first attempts to expli- citly apply Rosen’s (1974) theory of implicit markets to the development of a hedonic housing price model, included distance to the CBD among the neighbor- hood characteristics, and this loaded heavily on an JRAP 38(1):19-35. © 2008 MCRSA. All rights reserved. ===== PAGE 4 (score 30, numeric-paren 0, table-words 2, var-words 22) ===== 22 Ottensmann et al. Table 1. Descriptive Statistics Standard Variable Mean Deviation Sales price 110,888 78,492 Log sales price 11.437 0.621 Number of bathrooms 2.022 0.872 Total number of rooms 7.152 1.863 Floor space (100s sq ft) 16.371 7.592 Has basement 0.416 0.493 Age 37.276 26.882 Lot area <= 0.5 acre 0.843 0.364 Lot area >= 1.0 acre 0.033 0.179 Garage bays 1.622 0.742 Has porch or deck 0.562 0.496 Exterior brick or stone 0.603 0.489 No air conditioning 0.138 0.345 Effective tax rate 1.185 0.783 School district SAT 988.491 36.466 Neighborhood median income 50,837 18,166 Neighborhood percent black 20.736 21.746 Neighborhood percent vacant 7.529 4.161 Distance to CBD (kilometers) 11.710 4.880 Free-flow travel time to CBD 8.670 3.793 Congested travel time to CBD 8.630 3.937 Mean distance to 7 ZIP centers 15.459 3.923 Mean distance to 5 TAZ centers 15.840 4.190 Free-flow travel time to 5 TAZ cen- ters 12.937 3.329 Congested travel time to 5 TAZ centers 12.581 3.223 The remaining neighborhood characteristics, neighborhood median income, percentage of the pop- ulation African-American, and percentage of housing units vacant, have been derived from data from Cen- sus 2000 (U.S. Bureau of the Census 2006). Most often, census neighborhood characteristics have been in- cluded in hedonic models using values for the census al Progress (ISTEP), which is taken by all students. The SAT score variable was selected because it had a slightly higher level of signi- ficance in the model. tract or block group in which the housing unit is lo- cated. This approach has two problems. First, a house could be located near the boundary of a tract or block group so the relevant neighborhood characteristic is as much the characteristic of an adjacent area as the cha- racteristic of the area in which the house is located. Second, census tracts and block groups vary greatly in their spatial extent so census tract or block group measures represent characteristics extending at vary- ing distances from the housing unit. To address these problems in assigning census neighborhood characteristics, the model includes es- timates of these values for neighborhoods of a uniform one-mile radius surrounding each of the housing units. These neighborhood variables were estimated from the census block group data, with the values being the means of the census characteristics for the block groups intersecting the one-mile radius neigh- borhoods weighted by the proportion of the one-mile neighborhood area within each block group. Descrip- tive statistics for these neighborhood characteristics are included in Table 1. Several of the models tested include an additional neighborhood variable, location within Center Town- ship. Center Township is the older, inner-city portion of the urban area (see Figure 1 for the location). It is recognized locally as being a generally less-desirable area of the city. Figure 1. Location of Center Township ===== PAGE 5 (score 6, numeric-paren 0, table-words 0, var-words 6) ===== Location in a Hedonic Model 23 3.2 Employment Data The analyses test location measures related to the location of employment. Two sources of employment data are used, one with employment for the traffic analysis zones (TAZs) used for transportation plan- ning and a second with employment for ZIP code areas. Marion County is the inner part of the larger Indianapolis metropolitan area. Because accessibility to employment outside of Marion County is relevant to residents of the county, both sets of employment data include employment for Marion County and for the surrounding 8 counties. This area includes all of the counties in the Metropolitan Statistical Area (MSA) as defined in 2000 along with one additional county, Madison County, a separate MSA deemed potentially significant for employment accessibility. The first dataset has employment in 2000 for the 1285 TAZs in the 9-county area. These data were de- veloped from the ES-202 workforce data collected by the Indiana Department of Workforce Development. The data were assembled and provided by the Indiana Department of Transportation. The boundaries of the TAZs are shown in Figure 2. In addition to distance, the analyses also use the free-flow and congested road travel times between each pair of TAZs.2 These travel times were developed and provided by Cambridge Systematics, Inc. using a travel demand model origi- nally created by the Indianapolis Metropolitan Plan- ning Organization. These are travel times from the nodes within each TAZ at which trips to and from the TAZ are loaded onto the road network. The analyses involving travel times from houses to TAZ employ- ment use the travel times from the TAZ node closest to the house to the employment TAZ nodes. For the ana- lyses involving distance, the distances are the airline distances from the actual house locations to the em- ployment TAZ nodes. The second employment dataset includes em- ployment for both 1995 and 2000 for 100 ZIP code areas in the same 9-county area. These are also from the ES-202 workforce data collected by the Indiana Department of Workforce Development. The Indiana Business Research Center at IUPUI created the tabula- tions of these data by ZIP code for the Center for Ur- ban Policy and the Environment.3 The boundaries of 2 The free-flow travel times are the times for the shortest paths over the road network based upon the travel speeds on the network links under non-congested conditions. The congested travel times are the shortest paths taking into account the assignment of all peak period trips to the network, with the consequent reduction in travel speeds on some links as a result of the assigned trip volumes. 3 Employment for non-spatially-extensive ZIP codes such as post offices and individual firms and buildings (point locations) were the ZIP code areas are shown in Figure 3. Distances of house sales to ZIP code employment are the direct air- line distances from the house locations to point loca- tions within each ZIP code.4 Figure 2. TAZ Boundaries and Employment Centers The two employment datasets are used for several reasons. Only the ZIP code dataset includes employ- ment at two points in time, allowing examination of the effect of recent changes in accessibility to employ- ment. The travel time data correspond to the TAZ da- ta, requiring the use of those data for the analyses in- volving travel time. Comparison of the distance ana- lyses using the two datasets also provides for consid- eration of the effect of the spatial resolution of em- ployment data on the performance of measures of lo- cation. added to the employment for the ZIP code areas in which those points were located. 4 The point locations used for the ZIP codes are from the file of 5- digit ZIP points provided by ESRI with the ArcView software rather than the ZIP code centroids. Prior research has found that these points correspond more closely to the centers of the population and employment distribution within the ZIP code areas than the centro- ids. No documentation is provided, however, as to exactly how those point locations were established. ===== PAGE 8 (score 179, numeric-paren 139, table-words 2, var-words 22) ===== 26 Ottensmann et al. Table 2. Models with and without Distance to CBD and Center Township (t-values) Model 1 Model 2 Model 3 Model 4 Independent Base CBD Center Center + Variables Model Distance Township CBD Number of bathrooms 0.1104* 0.1103* 0.1132* 0.1121* (21.19) (21.14) (22.08) (21.88) Total number of rooms 0.0130* 0.0131* 0.0110* 0.0114* (6.15) (6.16) (5.28) (5.48) Floor space (100s sq ft) 0.0211* 0.0210* 0.0214* 0.0214* (34.06) (34.04) (35.23) (35.21) Has basement 0.0974* 0.0970* 0.1087* 0.1036* (15.44) (15.13) (17.43) (16.43) Age -0.0024* -0.0024* -0.0021* -0.0024* (15.71) (14.65) (14.17) (15.07) Lot area <= 0.5 acre -0.0492* -0.0493* -0.0416* -0.0426* (6.23) (6.24) (5.35) (5.49) Lot area >= 1.0 acre 0.0956* 0.0959* 0.0951* 0.0996* (6.10) (6.11) (6.18) (6.47) Has garage 0.0855* 0.0855* 0.0842* 0.0843* (19.81) (19.81) (19.83) (19.88) Has porch or deck 0.0477* 0.0478* 0.0456* 0.0470* (8.66) (8.67) (8.42) (8.68) Exterior brick or stone 0.0840* 0.0841* 0.0777* 0.0784* (13.95) (13.96) (13.12) (13.25) No air conditioning -0.2511* -0.2512* -0.2444* -0.2451* (29.12) (29.11) (28.80) (28.91) Effective tax rate -0.2393* -0.2393* -0.2385* -0.2376* (70.28) (70.20) (71.24) (71.01) School district SAT 0.0004* 0.0004* 0.0003* 0.0003* (5.56) (5.45) (4.60) (3.66) Neighborhood median income 0.0051* 0.0051* 0.0049* 0.0058* (22.36) (17.88) (22.15) (20.54) Neighborhood percent black -0.0005* -0.0005* -0.0009* -0.0010* (3.31) (3.31) (6.73) (6.99) Neighborhood percent vacant -0.0216* -0.0216* -0.0135* -0.0129* (22.34) (22.34) (12.74) (12.08) Center Township -0.1748* -0.1886* (17.50) (18.25) Distance to CBD -0.0004 -0.0055* (0.35) (5.10) R2 0.8579* 0.8579* 0.8627* 0.8631* *Significant at p<0.001 and without the Center Township dummy. The subs- tantive conclusions one would make regarding the performance of the alternative location measures are the same in nearly all instances. Since the focus of this paper is on the performance of the measures of loca- tion, the results without the additional measure of lo- cation provided by the Center Township variable are reported. ===== PAGE 9 (score 25, numeric-paren 0, table-words 6, var-words 6) ===== Location in a Hedonic Model 27 Travel times to the CBD provide alternative mo- nocentric measures of location that may more accu- rately represent accessibility to employment. Since 95 percent of workers in Marion County who commuted to work in 2000 used cars or trucks, and only 2.3 per- cent took public transit, travel time on the road net- work is the relevant measure for nearly all workers (U.S. Bureau of the Census 2006). Free-flow travel times are those estimated for conditions under which traffic volumes do not adversely affect speeds. Con- gested travel times are estimated for the peak period, taking into account traffic volumes that reduce speeds. Congested travel times thus may be more appropriate for accessibility to employment, with a high propor- tion of commuting taking place during the peak pe- riods. Results for distance, free-flow travel time, and congested travel time to the CBD are presented in the top portion of Table 3. (From this point forward, re- sults are presented only for the location variables and the overall R2 value for the model, as the regression coefficients for the other variables in the model are not significantly affected by the location measures.) As observed above, distance to the CBD was not signifi- cant. However, both free-flow and congested travel times to the CBD were significant and have the ex- pected negative signs for the regression coefficients. In addition, the regression coefficient and t value for congested travel time are twice as large as for free- flow travel times. So travel times to the CBD (especial- ly congested times) are significantly better predictors of house prices and better measures of accessibility to employment. Table 3. Models for Distance/Travel Time to CBD and Multiple Employment Centers Regression Standard Location Variable(s) Coefficient Error t Sig R2 Base Model 0.8579 Distance/Travel Time to CBD Distance to CBD -0.0004 0.0011 -0.3471 0.7285 0.8579 Free-flow travel time to CBD -0.0033 0.0013 -2.6385 0.0083 0.8580 Congested travel time to CBD -0.0066 0.0012 -5.2895 0.0000 0.8583 Distance/Travel Time to 3 TAZ Employment Centers Distance to CBD 0.0005 0.0012 0.4542 0.6497 0.8605 Distance to Keystone/Castleton -0.0008 0.0006 -1.2444 0.2134 Distance to West 86th Street -0.0048 0.0006 -7.9205 0.0000 Free-flow travel time to CBD -0.0032 0.0014 -2.3359 0.0195 0.8608 Free-flow time to Keyston/Castleton -0.0015 0.0006 -2.3661 0.0180 Free-flow time to West 86th Street -0.0060 0.0006 -9.2329 0.0000 Congested travel time to CBD -0.0050 0.0013 -3.8608 0.0001 0.8610 Congested time to Keystone/Castleton -0.0010 0.0006 -1.7612 0.0782 Congested time to West 86th Street -0.0063 0.0007 -9.4799 0.0000 Mean Distance/Travel Time to Employment Centers Mean distance to 7 ZIP centers -0.0096 0.0008 -11.6072 0.0000 0.8601 Mean distance to 5 TAZ centers -0.0089 0.0008 -11.8679 0.0000 0.8601 Mean free-flow time to 5 TAZ centers -0.0129 0.0010 -13.4577 0.0000 0.8608 Mean congested time to 5 TAZ centers -0.0138 0.0010 -14.0707 0.0000 0.8610 ===== PAGE 10 (score 12, numeric-paren 0, table-words 2, var-words 8) ===== 28 Ottensmann et al. 4.2. Distance/Travel Time to Employment Centers The increasing decentralization of employment within metropolitan areas has led to the questioning of the continued appropriateness of the monocentric as- sumption of the standard model. Observation of the emergence of multiple employment centers within metropolitan areas has led to suggestions that accessi- bility to employment in those multiple centers would provide a better measure of location within hedonic models. With distances or times to multiple employment centers, the issue arises as to how those centers would collectively influence house prices. Small and Song (1994) and Anas et al. (1998) lay out three alternatives (focusing on population densities rather than house prices). Assume that the effect of any employment center on house prices is an estimated house price val- ue at that center times a negative exponential function of distance or time from that center, as generally as- sumed with the monocentric model. In the first alter- native, if the centers are perfect substitutes, the effect on house prices is the maximum of the values pre- dicted for the various centers. In other words, each center independently affects house prices for those houses within its “sphere of influence.” At the other extreme, if the centers are complements, with the resi- dents of each house valuing access to every employ- ment center, the effect of the centers is the product of their individual effects. This has the very convenient property that when taking the log of the dependent variable, the effects of the distances or times are addi- tive, making the model easy to estimate. Finally, the third alternative represents an intermediate case, with the negative exponential effects of distance or time on house prices being additive. This method requires nonlinear estimation. It has been employed by Griffith (1981), Gordon et al. (1986), Song (1994), and Small and Song (1994) in predicting population densities but not within the more complex context of a hedonic housing price model. Most commonly used is the second formulation assuming that the centers are complements, undoubtedly due to the ease of estima- tion. This paper employs that approach, followed by a brief discussion of (unsuccessful) attempts to estimate the model for the other extreme, assuming the centers are substitutes. The obvious first attempts involved estimation of models for multiple employment centers by including the distances or times from the house sales locations to the all of the centers. The results for such models for both the TAZ and ZIP code centers were seriously compromised by multicollinearity. Regression coeffi- cient values and signs for the distance and time va- riables were virtually random. Variance Inflation Fac- tors were extremely high. The relatively close proximi- ty of the employment centers in Indianapolis appar- ently produced these levels of multicollinearity, as opposed to, for example, the more widely separated employment centers for which similar models were successfully estimated for Los Angeles (Heikkila et al. 1989). For the TAZ-based employment centers, three of the five centers identified had substantially greater employment than the other centers. These are the cen- ters shown darker and labeled in Figure 2. Models were estimated using the distances and times to just these three major employment centers. The results are shown in the middle of Table 3. Looking first at the overall performance of the models, the R2 values show increases over the base model, rising from 0.858 to about 0.860 to 0.861, as opposed to negligible increases for the models using only distances or times to the CBD. (It should be noted that the objective sought in considering the alternative measures of location was not, however, to substantially improve the predictive power of the model, which is already very high. Ra- ther, it was to provide for a more complete specifica- tion of the model.) As with distance to the CBD, the travel times performed better than distance, with con- gested travel times slightly better than free-flow times. For the two models using travel times, the regres- sion coefficients for the times to each of the three cen- ters were statistically significant. For the model using distance, only one of the coefficients, distance to Keys- tone/Castleton, was significant. Distance to the CBD and the other employment center were not. This is the one additional case in which differences are observed when the Center Township dummy variable was in- cluded in the model. With Center Township included to account for the negative neighborhood effects, the regression coefficients for both distance to the CBD and to the West 86th Street center were significant, paralleling the results obtained for the models using only distance to the CBD. The ZIP code-based centers outside of the CBD did not have the significant differences in total em- ployment levels observed for the TAZ centers. It was therefore not possible to identify a subset of major employment centers that could be used for estimating comparable models. Distances and times to all of the employment cen- ters can be included in models if one makes the as- sumption that the coefficients for the negative expo- nential decline of house prices are the same for each of the centers. With this assumption, models can be esti- ===== PAGE 12 (score 34, numeric-paren 0, table-words 12, var-words 5) ===== 30 Ottensmann et al. Table 4 Models for Accessibility to Employment and Change in Employment. Location Variable Regression Coefficient Standard Error t Sig R2 Accessibility to Employment Distance accessibility to ZIP employment (coefficient = 1.1) 0.0059 0.0006 10.1156 0.0000 0.8595 Distance accessibility to TAZ employment (coefficient = 1.6) 0.0205 0.0014 14.5388 0.0000 0.8612 Free-flow travel time accessibility to TAZ employment (coefficient = 2.2) 0.0273 0.0019 14.5371 0.0000 0.8612 Congested travel time accessibility to TAZ employment (coefficient = 2.1) 0.0232 0.0017 13.8811 0.0000 0.8610 Accessibility to Change in Employment Distance accessibility to change in ZIP employment (coefficient = 0.081) 0.0081 0.0006 13.5473 0.0000 0.8608 Accessibility to Employment and Change in Employment Distance accessibility to ZIP employment (coefficient = 1.3) 0.0048 0.0007 6.9897 0.0000 0.8616 Distance accessibility to change in ZIP employment (coefficient = 0.067) 0.0069 0.0006 11.4801 0.0000 Accessibility to Employment, Change in Employment, Mean Distance to ZIP Centers Distance accessibility to ZIP employment (coefficient = 1.3) 0.0048 0.0007 6.9896 0.0000 0.8616 Distance accessibility to change in ZIP employment (coefficient = 0.067) 0.0075 0.0012 6.0354 0.0000 Mean distance to 7 ZIP employment cen- ters 0.0008 0.0017 0.4736 0.6358 In his comparison of the performance of alterna- tive accessibility measures in predicting population densities, Song (1996) chose to use the natural log of accessibility to predict density rather than using the natural form as was done here. Tests were made using the log of the accessibilities, and the fits of the models (after finding best-fit accessibility coefficients asso- ciated with using the logs) were no better. 4.4. Change in Accessibility to Employment All of the applications involving the use of meas- ures of accessibility to employment to predict densities or house prices have used accessibility to levels of em- ployment at a single point in time. Attention has not been given to the possible effect of changes in accessi- bility to employment over time. However, if accessibil- ity to employment has a significant effect on house prices, then changes in accessibility to employment in the recent past, which may be seen as being predictive of future changes in accessibility, might also affect house prices. A study examining changes in accessibil- ity to employment in Milwaukee over more than three decades found these changes were associated with a range of neighborhood characteristics (Ottensmann 1980). In the estimation of predictive models for an urban simulation model, change in accessibility to employment was included along with current accessi- bility in the prediction of the change from nonurban to residential land use (Ottensmann 2005). ===== PAGE 13 (score 32, numeric-paren 0, table-words 10, var-words 12) ===== Location in a Hedonic Model 31 A model was estimated examining the effect on house prices in the hedonic model of changes in acces- sibility to employment from 1995 to 2000 using the distances to ZIP code employment. The assumption is made that the accessibility coefficient is the same in both years, with the calculation actually being of ac- cessibility to change in employment, which is mathe- matically equivalent using this assumption. The best- fit accessibility coefficient was determined as for ac- cessibility to employment. The results are shown in the next portion of Table 4. The fit of the model was about as good as for the various accessibility measures (and even better than for accessibility to ZIP code em- ployment). Accessibility to employment change was statistically significant. If both accessibility to employment and change in accessibility to employment are effective measures of location for use in the prediction of house prices, it is not unreasonable to consider that the combination of the two might result in even better predictions. A model including both accessibility measures was esti- mated, with the results presented in the next section of Table 4. Estimation began by using the best-fit accessi- bility coefficients estimated for accessibility and change in accessibility when used separately. Both coefficients were then adjusted iteratively until the combination of accessibility coefficients producing the lowest standard error of estimate was found.7 This model has an R2 value of 0.862, which is higher than for any of the other models examined. The regression coefficients for both accessibility to employment and change in accessibility to employment are statistically significant, with the latter having the higher t value. So change in accessibility to employment is a significant predictor of house prices after controlling for the ef- fects of the level of accessibility to employment. To- gether, the two provide the best measure of location for predicting house prices. With both accessibility to employment and change in accessibility significant predictors of house prices, the question arises as to whether distances to the mul- tiple employment centers would also contribute in measuring location. The mean distance to the ZIP em- ployment centers was added to the previous model, and the results are shown in the final section of Table 4. Mean distance to the employment centers was not 7 In searching for the best combination of accessibility coefficients, both the high and low estimates of the accessibility to employment coefficients as described in the previous footnote were used as start- ing points. Once again, different combinations of accessibility coeffi- cients yielded two different minima with nearly identical standard errors. One had a negative regression coefficient for accessibility. Results for the model in which both coefficients were positive, as expected, are reported. statistically significant, and the regression coefficients for the accessibility measures changed little. The com- bination of the two accessibility measures appears to capture the effect of employment location, with the location of multiple employment centers adding noth- ing. 5. Conclusions The measures of location were statistically signifi- cant predictors of sales price in the hedonic model. However, their addition to the model produced only small increases in R2. This raises the question as to the relative importance of location as opposed to house and neighborhood characteristics in determining house prices. Table 5 presents the predicted changes in sales price for changes in the independent variables for a house with the mean sales price of $110,888. Changes for the house and neighborhood variables have been calculated using the regression coefficients estimated for the base model without location. For the changes in distances and times, changes are shown for 10 kilometer and 10 minute changes in those meas- ures. Changes for the house and neighborhood meas- ures are for unit changes in those measures, with the exception of the effective tax rate and the school dis- trict SAT, for which more intuitive levels of change were used. As expected, the house characteristics are major determinants of sales price. Unit changes in four of the variables—number of bathrooms, lot size greater than one acre, and presence of a basement and air condi- tioning—resulted in changes in sales price greater than 10 percent. The effects of the neighborhood variables were smaller, though the percentage of units in the neighborhood vacant and, especially, school district SAT scores, had substantial effects. Considering distance and travel time to the CBD, the effect of distance was small, but this variable was not significant in the model. Travel times to the CBD had a much larger effect, with 10 minute increases in travel times being associated with 3.3 and 6.4 percent declines in sales price for the free-flow and congested travel time models respectively. Mean distance and travel times to the multiple employment centers had greater effects on sales price, with 10 kilometer or 10 minutes increases in the means predicting 9.1 to 12.9 percent decreases in sales prices, decreases in prices for a house with the mean price ranging from $9,500 to $12,900. These are substantial effects of location with respect to employment on house prices. ===== PAGE 14 (score 32, numeric-paren 0, table-words 2, var-words 26) ===== 32 Ottensmann et al. Table 5. Change in Sales Price at Mean Sales Price Sales Price Percent Dependent Variable Change Change Change Number of bathrooms $12,945 11.67 Total number of rooms $1,453 1.31 Floor space (100 sq ft) $2,359 2.13 Has basement $11,347 10.23 Age -$263 -0.24 Lot area <= 0.5 acre -$5,321 -4.80 Lot area >= 1.0 acre $11,119 10.03 Garage bays $9,902 8.93 Has porch or deck $5,420 4.89 Exterior brick or stone $9,714 8.76 No air conditioning -$24,626 -22.21 Effective tax rate (tax per $100) -$2,622 -2.36 School district SAT (100 points) $4,592 4.14 Neighborhood median income $564 0.51 Neighborhood percent black -$51 -0.05 Neighborhood percent vacant -$2,373 -2.14 Distance to CBD (10 kilometers) -$408 -0.37 Free-flow travel time to CBD (10 minutes) -$3,642 -3.28 Congested travel time to CBD (10 minutes) -$7,039 -6.35 Mean distance to 7 ZIP centers (10 kilometers) -$10,144 -9.15 Mean distance to 5 TAZ centers (10 kilometers) -$9,483 -8.55 Mean free-flow travel time to 5 TAZ centers (10 minutes) -$13,426 -12.11 Mean congested travel time to 5 TAZ centers (10 minutes) -$14,307 -12.90 Distance and travel times to the CBD were the least effective measures of location in predicting house prices in the hedonic model. Distance to the CBD was not even statistically significant in the base model. It was, however, significant when location in Center Township (the inner city) was included in the model to control for externalities negatively affecting house prices. These results may provide some insight into the frequent results of insignificant or even positive coefficients for distance to the CBD in hedonic models (Bender and Hwang 1985; Coulson 1991). If there are negative neighborhood effects in inner-city areas that are not completely reflected in the neighborhood cha- racteristics included in a hedonic model, effects that are related to distance to the center, the finding of no significant negative decline of house price with dis- tance may reflect incomplete specification of the neighborhood effects in the model. This suggests cau- tion is in order in the interpretation of findings of a lack of significant decline in house prices with distance from the CBD. For example, Heikkila et al. (1991) em- phasized the result that, in their hedonic model in- cluding distances to multiple employment centers in Los Angeles, distance to the CBD is the one measure that was not statistically significant. While it certainly may be true that the Los Angeles CBD does not affect house prices in the area, it is also possible that the fail- ure to observe an effect resulted from an inadequate specification of neighborhood effects that are asso- ciated with distance to the CBD. This finding empha- sizes the importance of correctly specifying neighbor- hood effects in hedonic models. Measures of location including employment out- side of the CBD yielded better predictions and more significant measures of location. Measures of distance or time to multiple employment centers performed about as well as the measures of accessibility to em- ployment. These results do not show clearly better performance for the accessibility measures as found by Song (1994; 1996) in models predicting population density. Across the models using distances or times to the CBD and to multiple employment centers, travel times consistently performed better than distances. Among the alternative travel time measures, congested travel times did better than free-flow travel times. This pat- tern did not hold, however, for the accessibility to em- ployment measures. For accessibility to employment, the measures of accessibility to the more spatially- disaggregated TAZ employment performed better than accessibility to employment in the ZIP code areas. Recent change in accessibility to employment proved to be a significant measure of location affecting house prices. This variable, calculated using ZIP code employment change, actually performed better than accessibility to ZIP code area employment. The com- bination of accessibility to employment with change in accessibility was the best overall measure of location in predicting house prices, with both measures signifi- cant. Location with respect to employment is a signifi- cant predictor of house prices within a hedonic model, supporting the general assumption embedded in the standard urban economic model. With the decentrali- ===== PAGE 15 (score 9, numeric-paren 0, table-words 0, var-words 9) ===== Location in a Hedonic Model 33 zation of employment that has occurred, the findings do not, however, support the specific assumption of the standard monocentric model of employment lo- cated only in the CBD. Measures of location that in- clude employment outside of the CBD do better than the simple measures of distance or time to the CBD. These results are not only important for under- standing the nature of the effects of accessibility to employment on house prices. They also have impor- tant implications for public policy. Most obvious is the use of hedonic house price models and other types of models for property assessment for tax purposes. In- creasing use is being made of such models for the ap- praisal of real property and for real estate assessment (see, e.g., Detweiler and Radigan 1996; Pagourtzi et al. 2003; Renne 2003). Obtaining accurate assessment de- pends upon the proper specification of the models used. Because location with respect to employment has a significant effect on house values, failure to in- clude appropriate measures of such accessibility can result in significant errors in assessment. Hedonic house price models are frequently em- ployed to estimate the benefits or costs associated with location-based amenities or disamenities, often with the goal of informing public policymaking to address the amenities or disamenities. It is a truism that the misspecification of a model can result in errors in the estimation of the model coefficients, though often the direction and magnitude of such errors is unknown. However, in at least some applications of hedonic models to estimate the negative effects of disamenities, the effect of failure to properly specify location with respect to employment can be quite clear. Consider the case of the effect of airport noise on property values. (See Nelson 2004 for references to a large number of such studies.) Areas near airports are frequently major employment centers. If accessibility to that employ- ment is not included in the hedonic model, predicted house prices near the airport will be underestimated. This may then result in an underestimate of the nega- tive effect of airport noise on those properties affected. Two studies (Espey and Lopez 2000; Tomkins et al. 1998) have used hedonic models to simultaneously estimate the effects on property values of both airport noise and proximity to the airport. Both studies consi- dered the airport as a potential amenity, not as an em- ployment center, and no measures of location with respect to employment were included in either model. Both found negative effects of airport noise on proper- ty values, but they differed as to whether proximity had a positive or negative effect. Tomkins et al. (1998) found proximity to the Manchester airport led to high- er values. Espey and Lopez (2000) found a negative effect for the Reno-Tahoe airport, but they noted that the area surrounding that airport is primarily residen- tial, so it may not be a significant employment center. Another example where disamenities might be asso- ciated with employment concentrations would be air pollution and certain types of manufacturing em- ployment. Hedonic models using measures of accessibility based on travel times provide an opportunity to esti- mate the value people place on travel time savings. Such estimates can be used in valuing the benefits as- sociated with travel time reductions associated with planned transportation system improvements. At least one author (Nelson 1977) has directly employed a he- donic house price model for this purpose. Because it is a significant predictor of house prices, location with respect to employment should be in- cluded for the proper specification of hedonic housing price models. The combination of accessibility to em- ployment and change in accessibility to employment provides the best specification of location, but the ne- cessary data may not always be available. But certain- ly location with respect to employment should include either distances or times to multiple employment cen- ters or a measure of accessibility to all employment, as the assumption that access to employment in the CBD would be sufficient is not supported. References Adair, A., S. McGreal, A. Smyth, J. Cooper and T. Ry- ley. 2000. House prices and accessibility: The test- ing of relationships within the Belfast urban area. Housing Studies 15(5): 699-716. Alonso, W. 1964. Location and Land Use: Toward a Gen- eral Theory of Land Rent. Cambridge, MA: Harvard University Press. Anas, A., R. Arnott and K. Small. 1998. Urban spatial structure. Journal of Economic Literature 36(3): 1426- 1464. Bender, B. and H. Hwang. 1985. Hedonic house price indices and secondary employment centers. Journal of Urban Economics 17(1): 90-107. Bowen, W. M., B. A. Mikelbank and D. M. Preste- gaard. 2001. Theoretical and empirical considera- tions regarding space in hedonic housing price model applications. Growth and Change 32(4): 466- 490. Burnell, J. D. 1985. Industrial land use, externalities, and residential location. Urban Studies 22(5): 399- 408. Coulson, N. E. 1991. Really useful tests of the mono- centric model. Land Economics 67(3): 299-307.